git.sesse.net Git - nageru/blob - nageru/mixer.cpp

   1 #undef Success
   2
   3 #include "mixer.h"
   4
   5 #include <assert.h>
   6 #include <epoxy/egl.h>
   7 #include <movit/effect.h>
   8 #include <movit/effect_chain.h>
   9 #include <movit/effect_util.h>
  10 #include <movit/flat_input.h>
  11 #include <movit/image_format.h>
  12 #include <movit/init.h>
  13 #include <movit/resource_pool.h>
  14 #include <pthread.h>
  15 #include <stdint.h>
  16 #include <stdio.h>
  17 #include <stdlib.h>
  18 #include <algorithm>
  19 #include <chrono>
  20 #include <condition_variable>
  21 #include <cstddef>
  22 #include <cstdint>
  23 #include <memory>
  24 #include <mutex>
  25 #include <ratio>
  26 #include <string>
  27 #include <thread>
  28 #include <utility>
  29 #include <vector>
  30
  31 #include "DeckLinkAPI.h"
  32 #include "LinuxCOM.h"
  33 #include "alsa_output.h"
  34 #include "basic_stats.h"
  35 #include "bmusb/bmusb.h"
  36 #include "bmusb/fake_capture.h"
  37 #ifdef HAVE_CEF
  38 #include "cef_capture.h"
  39 #endif
  40 #include "chroma_subsampler.h"
  41 #include "shared/context.h"
  42 #include "decklink_capture.h"
  43 #include "decklink_output.h"
  44 #include "defs.h"
  45 #include "shared/disk_space_estimator.h"
  46 #include "ffmpeg_capture.h"
  47 #include "flags.h"
  48 #include "image_input.h"
  49 #include "input_mapping.h"
  50 #include "shared/metrics.h"
  51 #include "mjpeg_encoder.h"
  52 #include "pbo_frame_allocator.h"
  53 #include "shared/ref_counted_gl_sync.h"
  54 #include "resampling_queue.h"
  55 #include "shared/timebase.h"
  56 #include "timecode_renderer.h"
  57 #include "v210_converter.h"
  58 #include "va_display_with_cleanup.h"
  59 #include "video_encoder.h"
  60
  61 #undef Status
  62 #include <google/protobuf/util/json_util.h>
  63 #include "json.pb.h"
  64
  65 class IDeckLink;
  66 class QOpenGLContext;
  67
  68 using namespace movit;
  69 using namespace std;
  70 using namespace std::chrono;
  71 using namespace std::placeholders;
  72 using namespace bmusb;
  73
  74 Mixer *global_mixer = nullptr;
  75
  76 namespace {
  77
  78 void insert_new_frame(RefCountedFrame frame, unsigned field_num, bool interlaced, unsigned card_index, InputState *input_state)
  79 {
  80         if (interlaced) {
  81                 for (unsigned frame_num = FRAME_HISTORY_LENGTH; frame_num --> 1; ) {  // :-)
  82                         input_state->buffered_frames[card_index][frame_num] =
  83                                 input_state->buffered_frames[card_index][frame_num - 1];
  84                 }
  85                 input_state->buffered_frames[card_index][0] = { frame, field_num };
  86         } else {
  87                 for (unsigned frame_num = 0; frame_num < FRAME_HISTORY_LENGTH; ++frame_num) {
  88                         input_state->buffered_frames[card_index][frame_num] = { frame, field_num };
  89                 }
  90         }
  91 }
  92
  93 void ensure_texture_resolution(PBOFrameAllocator::Userdata *userdata, unsigned field, unsigned width, unsigned height, unsigned cbcr_width, unsigned cbcr_height, unsigned v210_width)
  94 {
  95         bool first;
  96         switch (userdata->pixel_format) {
  97         case PixelFormat_10BitYCbCr:
  98                 first = userdata->tex_v210[field] == 0 || userdata->tex_444[field] == 0;
  99                 break;
 100         case PixelFormat_8BitYCbCr:
 101                 first = userdata->tex_y[field] == 0 || userdata->tex_cbcr[field] == 0;
 102                 break;
 103         case PixelFormat_8BitBGRA:
 104                 first = userdata->tex_rgba[field] == 0;
 105                 break;
 106         case PixelFormat_8BitYCbCrPlanar:
 107                 first = userdata->tex_y[field] == 0 || userdata->tex_cb[field] == 0 || userdata->tex_cr[field] == 0;
 108                 break;
 109         default:
 110                 assert(false);
 111         }
 112
 113         const bool recreate_main_texture =
 114                 first ||
 115                 width != userdata->last_width[field] ||
 116                 height != userdata->last_height[field] ||
 117                 cbcr_width != userdata->last_cbcr_width[field] ||
 118                 cbcr_height != userdata->last_cbcr_height[field];
 119         const bool recreate_v210_texture =
 120                 global_flags.ten_bit_input &&
 121                 (first || v210_width != userdata->last_v210_width[field] || height != userdata->last_height[field]);
 122
 123         if (recreate_main_texture) {
 124                 // We changed resolution since last use of this texture, so we need to create
 125                 // a new object. Note that this each card has its own PBOFrameAllocator,
 126                 // we don't need to worry about these flip-flopping between resolutions.
 127                 switch (userdata->pixel_format) {
 128                 case PixelFormat_10BitYCbCr:
 129                         glBindTexture(GL_TEXTURE_2D, userdata->tex_444[field]);
 130                         check_error();
 131                         glTexImage2D(GL_TEXTURE_2D, 0, GL_RGB10_A2, width, height, 0, GL_RGBA, GL_UNSIGNED_INT_2_10_10_10_REV, nullptr);
 132                         check_error();
 133                         break;
 134                 case PixelFormat_8BitYCbCr: {
 135                         glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
 136                         check_error();
 137                         glTexImage2D(GL_TEXTURE_2D, 0, GL_RG8, cbcr_width, height, 0, GL_RG, GL_UNSIGNED_BYTE, nullptr);
 138                         check_error();
 139                         glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
 140                         check_error();
 141                         glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, width, height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
 142                         check_error();
 143                         break;
 144                 }
 145                 case PixelFormat_8BitYCbCrPlanar: {
 146                         glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
 147                         check_error();
 148                         glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, width, height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
 149                         check_error();
 150                         glBindTexture(GL_TEXTURE_2D, userdata->tex_cb[field]);
 151                         check_error();
 152                         glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, cbcr_width, cbcr_height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
 153                         check_error();
 154                         glBindTexture(GL_TEXTURE_2D, userdata->tex_cr[field]);
 155                         check_error();
 156                         glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, cbcr_width, cbcr_height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
 157                         check_error();
 158                         break;
 159                 }
 160                 case PixelFormat_8BitBGRA:
 161                         glBindTexture(GL_TEXTURE_2D, userdata->tex_rgba[field]);
 162                         check_error();
 163                         // NOTE: sRGB may be disabled by sRGBSwitchingFlatInput.
 164                         glTexImage2D(GL_TEXTURE_2D, 0, GL_SRGB8_ALPHA8, width, height, 0, GL_BGRA, GL_UNSIGNED_INT_8_8_8_8_REV, nullptr);
 165                         check_error();
 166                         break;
 167                 default:
 168                         assert(false);
 169                 }
 170                 userdata->last_width[field] = width;
 171                 userdata->last_height[field] = height;
 172                 userdata->last_cbcr_width[field] = cbcr_width;
 173                 userdata->last_cbcr_height[field] = cbcr_height;
 174         }
 175         if (recreate_v210_texture) {
 176                 // Same as above; we need to recreate the texture.
 177                 glBindTexture(GL_TEXTURE_2D, userdata->tex_v210[field]);
 178                 check_error();
 179                 glTexImage2D(GL_TEXTURE_2D, 0, GL_RGB10_A2, v210_width, height, 0, GL_RGBA, GL_UNSIGNED_INT_2_10_10_10_REV, nullptr);
 180                 check_error();
 181                 userdata->last_v210_width[field] = v210_width;
 182                 userdata->last_height[field] = height;
 183         }
 184 }
 185
 186 void upload_texture(GLuint tex, GLuint width, GLuint height, GLuint stride, bool interlaced_stride, GLenum format, GLenum type, GLintptr offset)
 187 {
 188         if (interlaced_stride) {
 189                 stride *= 2;
 190         }
 191         if (global_flags.flush_pbos) {
 192                 glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, offset, stride * height);
 193                 check_error();
 194         }
 195
 196         glBindTexture(GL_TEXTURE_2D, tex);
 197         check_error();
 198         if (interlaced_stride) {
 199                 glPixelStorei(GL_UNPACK_ROW_LENGTH, width * 2);
 200                 check_error();
 201         } else {
 202                 glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
 203                 check_error();
 204         }
 205
 206         glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, width, height, format, type, BUFFER_OFFSET(offset));
 207         check_error();
 208         glBindTexture(GL_TEXTURE_2D, 0);
 209         check_error();
 210         glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
 211         check_error();
 212 }
 213
 214 }  // namespace
 215
 216 void JitterHistory::register_metrics(const vector<pair<string, string>> &labels)
 217 {
 218         global_metrics.add("input_underestimated_jitter_frames", labels, &metric_input_underestimated_jitter_frames);
 219         global_metrics.add("input_estimated_max_jitter_seconds", labels, &metric_input_estimated_max_jitter_seconds, Metrics::TYPE_GAUGE);
 220 }
 221
 222 void JitterHistory::unregister_metrics(const vector<pair<string, string>> &labels)
 223 {
 224         global_metrics.remove("input_underestimated_jitter_frames", labels);
 225         global_metrics.remove("input_estimated_max_jitter_seconds", labels);
 226 }
 227
 228 void JitterHistory::frame_arrived(steady_clock::time_point now, int64_t frame_duration, size_t dropped_frames)
 229 {
 230         if (expected_timestamp > steady_clock::time_point::min()) {
 231                 expected_timestamp += dropped_frames * nanoseconds(frame_duration * 1000000000 / TIMEBASE);
 232                 double jitter_seconds = fabs(duration<double>(expected_timestamp - now).count());
 233                 history.push_back(orders.insert(jitter_seconds));
 234                 if (jitter_seconds > estimate_max_jitter()) {
 235                         ++metric_input_underestimated_jitter_frames;
 236                 }
 237
 238                 metric_input_estimated_max_jitter_seconds = estimate_max_jitter();
 239
 240                 if (history.size() > history_length) {
 241                         orders.erase(history.front());
 242                         history.pop_front();
 243                 }
 244                 assert(history.size() <= history_length);
 245         }
 246         expected_timestamp = now + nanoseconds(frame_duration * 1000000000 / TIMEBASE);
 247 }
 248
 249 double JitterHistory::estimate_max_jitter() const
 250 {
 251         if (orders.empty()) {
 252                 return 0.0;
 253         }
 254         size_t elem_idx = lrint((orders.size() - 1) * percentile);
 255         if (percentile <= 0.5) {
 256                 return *next(orders.begin(), elem_idx) * multiplier;
 257         } else {
 258                 return *prev(orders.end(), orders.size() - elem_idx) * multiplier;
 259         }
 260 }
 261
 262 void QueueLengthPolicy::register_metrics(const vector<pair<string, string>> &labels)
 263 {
 264         global_metrics.add("input_queue_safe_length_frames", labels, &metric_input_queue_safe_length_frames, Metrics::TYPE_GAUGE);
 265 }
 266
 267 void QueueLengthPolicy::unregister_metrics(const vector<pair<string, string>> &labels)
 268 {
 269         global_metrics.remove("input_queue_safe_length_frames", labels);
 270 }
 271
 272 void QueueLengthPolicy::update_policy(steady_clock::time_point now,
 273                                       steady_clock::time_point expected_next_frame,
 274                                       int64_t input_frame_duration,
 275                                       int64_t master_frame_duration,
 276                                       double max_input_card_jitter_seconds,
 277                                       double max_master_card_jitter_seconds)
 278 {
 279         double input_frame_duration_seconds = input_frame_duration / double(TIMEBASE);
 280         double master_frame_duration_seconds = master_frame_duration / double(TIMEBASE);
 281
 282         // Figure out when we can expect the next frame for this card, assuming
 283         // worst-case jitter (ie., the frame is maximally late).
 284         double seconds_until_next_frame = max(duration<double>(expected_next_frame - now).count() + max_input_card_jitter_seconds, 0.0);
 285
 286         // How many times are the master card expected to tick in that time?
 287         // We assume the master clock has worst-case jitter but not any rate
 288         // discrepancy, ie., it ticks as early as possible every time, but not
 289         // cumulatively.
 290         double frames_needed = (seconds_until_next_frame + max_master_card_jitter_seconds) / master_frame_duration_seconds;
 291
 292         // As a special case, if the master card ticks faster than the input card,
 293         // we expect the queue to drain by itself even without dropping. But if
 294         // the difference is small (e.g. 60 Hz master and 59.94 input), it would
 295         // go slowly enough that the effect wouldn't really be appreciable.
 296         // We account for this by looking at the situation five frames ahead,
 297         // assuming everything else is the same.
 298         double frames_allowed;
 299         if (master_frame_duration < input_frame_duration) {
 300                 frames_allowed = frames_needed + 5 * (input_frame_duration_seconds - master_frame_duration_seconds) / master_frame_duration_seconds;
 301         } else {
 302                 frames_allowed = frames_needed;
 303         }
 304
 305         safe_queue_length = max<int>(floor(frames_allowed), 0);
 306         metric_input_queue_safe_length_frames = safe_queue_length;
 307 }
 308
 309 Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards)
 310         : httpd(),
 311           num_cards(num_cards),
 312           mixer_surface(create_surface(format)),
 313           h264_encoder_surface(create_surface(format)),
 314           decklink_output_surface(create_surface(format)),
 315           image_update_surface(create_surface(format))
 316 {
 317         memcpy(ycbcr_interpretation, global_flags.ycbcr_interpretation, sizeof(ycbcr_interpretation));
 318         CHECK(init_movit(MOVIT_SHADER_DIR, MOVIT_DEBUG_OFF));
 319         check_error();
 320
 321         if (!epoxy_has_gl_extension("GL_EXT_texture_sRGB_decode") ||
 322             !epoxy_has_gl_extension("GL_ARB_sampler_objects")) {
 323                 fprintf(stderr, "Nageru requires GL_EXT_texture_sRGB_decode and GL_ARB_sampler_objects to run.\n");
 324                 exit(1);
 325         }
 326
 327         // Since we allow non-bouncing 4:2:2 YCbCrInputs, effective subpixel precision
 328         // will be halved when sampling them, and we need to compensate here.
 329         movit_texel_subpixel_precision /= 2.0;
 330
 331         resource_pool.reset(new ResourcePool);
 332         for (unsigned i = 0; i < NUM_OUTPUTS; ++i) {
 333                 output_channel[i].parent = this;
 334                 output_channel[i].channel = i;
 335         }
 336
 337         ImageFormat inout_format;
 338         inout_format.color_space = COLORSPACE_sRGB;
 339         inout_format.gamma_curve = GAMMA_sRGB;
 340
 341         // Matches the 4:2:0 format created by the main chain.
 342         YCbCrFormat ycbcr_format;
 343         ycbcr_format.chroma_subsampling_x = 2;
 344         ycbcr_format.chroma_subsampling_y = 2;
 345         if (global_flags.ycbcr_rec709_coefficients) {
 346                 ycbcr_format.luma_coefficients = YCBCR_REC_709;
 347         } else {
 348                 ycbcr_format.luma_coefficients = YCBCR_REC_601;
 349         }
 350         ycbcr_format.full_range = false;
 351         ycbcr_format.num_levels = 1 << global_flags.x264_bit_depth;
 352         ycbcr_format.cb_x_position = 0.0f;
 353         ycbcr_format.cr_x_position = 0.0f;
 354         ycbcr_format.cb_y_position = 0.5f;
 355         ycbcr_format.cr_y_position = 0.5f;
 356
 357         // Initialize the neutral colors to sane values.
 358         for (unsigned i = 0; i < MAX_VIDEO_CARDS; ++i) {
 359                 last_received_neutral_color[i] = RGBTriplet(1.0f, 1.0f, 1.0f);
 360         }
 361
 362         // Display chain; shows the live output produced by the main chain (or rather, a copy of it).
 363         display_chain.reset(new EffectChain(global_flags.width, global_flags.height, resource_pool.get()));
 364         check_error();
 365         GLenum type = global_flags.x264_bit_depth > 8 ? GL_UNSIGNED_SHORT : GL_UNSIGNED_BYTE;
 366         display_input = new YCbCrInput(inout_format, ycbcr_format, global_flags.width, global_flags.height, YCBCR_INPUT_SPLIT_Y_AND_CBCR, type);
 367         display_chain->add_input(display_input);
 368         display_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
 369         display_chain->set_dither_bits(0);  // Don't bother.
 370         display_chain->finalize();
 371
 372         video_encoder.reset(new VideoEncoder(resource_pool.get(), h264_encoder_surface, global_flags.va_display, global_flags.width, global_flags.height, &httpd, global_disk_space_estimator));
 373         if (!global_flags.card_to_mjpeg_stream_export.empty()) {
 374                 mjpeg_encoder.reset(new MJPEGEncoder(&httpd, global_flags.va_display));
 375         }
 376
 377         // Must be instantiated after VideoEncoder has initialized global_flags.use_zerocopy.
 378         theme.reset(new Theme(global_flags.theme_filename, global_flags.theme_dirs, resource_pool.get(), num_cards));
 379
 380         // Must be instantiated after the theme, as the theme decides the number of FFmpeg inputs.
 381         std::vector<FFmpegCapture *> video_inputs = theme->get_video_inputs();
 382         audio_mixer.reset(new AudioMixer(num_cards, video_inputs.size()));
 383
 384         httpd.add_endpoint("/channels", bind(&Mixer::get_channels_json, this), HTTPD::ALLOW_ALL_ORIGINS);
 385         for (int channel_idx = 0; channel_idx < theme->get_num_channels(); ++channel_idx) {
 386                 char url[256];
 387                 snprintf(url, sizeof(url), "/channels/%d/color", channel_idx + 2);
 388                 httpd.add_endpoint(url, bind(&Mixer::get_channel_color_http, this, unsigned(channel_idx + 2)), HTTPD::ALLOW_ALL_ORIGINS);
 389         }
 390
 391         // Start listening for clients only once VideoEncoder has written its header, if any.
 392         httpd.start(global_flags.http_port);
 393
 394         // First try initializing the then PCI devices, then USB, then
 395         // fill up with fake cards until we have the desired number of cards.
 396         unsigned num_pci_devices = 0;
 397         unsigned card_index = 0;
 398
 399         {
 400                 IDeckLinkIterator *decklink_iterator = CreateDeckLinkIteratorInstance();
 401                 if (decklink_iterator != nullptr) {
 402                         for ( ; card_index < num_cards; ++card_index) {
 403                                 IDeckLink *decklink;
 404                                 if (decklink_iterator->Next(&decklink) != S_OK) {
 405                                         break;
 406                                 }
 407
 408                                 DeckLinkCapture *capture = new DeckLinkCapture(decklink, card_index);
 409                                 DeckLinkOutput *output = new DeckLinkOutput(resource_pool.get(), decklink_output_surface, global_flags.width, global_flags.height, card_index);
 410                                 if (!output->set_device(decklink)) {
 411                                         delete output;
 412                                         output = nullptr;
 413                                 }
 414                                 configure_card(card_index, capture, CardType::LIVE_CARD, output);
 415                                 ++num_pci_devices;
 416                         }
 417                         decklink_iterator->Release();
 418                         fprintf(stderr, "Found %u DeckLink PCI card(s).\n", num_pci_devices);
 419                 } else {
 420                         fprintf(stderr, "DeckLink drivers not found. Probing for USB cards only.\n");
 421                 }
 422         }
 423
 424         unsigned num_usb_devices = BMUSBCapture::num_cards();
 425         for (unsigned usb_card_index = 0; usb_card_index < num_usb_devices && card_index < num_cards; ++usb_card_index, ++card_index) {
 426                 BMUSBCapture *capture = new BMUSBCapture(usb_card_index);
 427                 capture->set_card_disconnected_callback(bind(&Mixer::bm_hotplug_remove, this, card_index));
 428                 configure_card(card_index, capture, CardType::LIVE_CARD, /*output=*/nullptr);
 429         }
 430         fprintf(stderr, "Found %u USB card(s).\n", num_usb_devices);
 431
 432         unsigned num_fake_cards = 0;
 433         for ( ; card_index < num_cards; ++card_index, ++num_fake_cards) {
 434                 FakeCapture *capture = new FakeCapture(global_flags.width, global_flags.height, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
 435                 configure_card(card_index, capture, CardType::FAKE_CAPTURE, /*output=*/nullptr);
 436         }
 437
 438         if (num_fake_cards > 0) {
 439                 fprintf(stderr, "Initialized %u fake cards.\n", num_fake_cards);
 440         }
 441
 442         // Initialize all video inputs the theme asked for. Note that these are
 443         // all put _after_ the regular cards, which stop at <num_cards> - 1.
 444         for (unsigned video_card_index = 0; video_card_index < video_inputs.size(); ++card_index, ++video_card_index) {
 445                 if (card_index >= MAX_VIDEO_CARDS) {
 446                         fprintf(stderr, "ERROR: Not enough card slots available for the videos the theme requested.\n");
 447                         abort();
 448                 }
 449                 configure_card(card_index, video_inputs[video_card_index], CardType::FFMPEG_INPUT, /*output=*/nullptr);
 450                 video_inputs[video_card_index]->set_card_index(card_index);
 451         }
 452         num_video_inputs = video_inputs.size();
 453
 454 #ifdef HAVE_CEF
 455         // Same, for HTML inputs.
 456         std::vector<CEFCapture *> html_inputs = theme->get_html_inputs();
 457         for (unsigned html_card_index = 0; html_card_index < html_inputs.size(); ++card_index, ++html_card_index) {
 458                 if (card_index >= MAX_VIDEO_CARDS) {
 459                         fprintf(stderr, "ERROR: Not enough card slots available for the HTML inputs the theme requested.\n");
 460                         abort();
 461                 }
 462                 configure_card(card_index, html_inputs[html_card_index], CardType::CEF_INPUT, /*output=*/nullptr);
 463                 html_inputs[html_card_index]->set_card_index(card_index);
 464         }
 465         num_html_inputs = html_inputs.size();
 466 #endif
 467
 468         BMUSBCapture::set_card_connected_callback(bind(&Mixer::bm_hotplug_add, this, _1));
 469         BMUSBCapture::start_bm_thread();
 470
 471         for (unsigned card_index = 0; card_index < num_cards + num_video_inputs + num_html_inputs; ++card_index) {
 472                 cards[card_index].queue_length_policy.reset(card_index);
 473         }
 474
 475         chroma_subsampler.reset(new ChromaSubsampler(resource_pool.get()));
 476
 477         if (global_flags.ten_bit_input) {
 478                 if (!v210Converter::has_hardware_support()) {
 479                         fprintf(stderr, "ERROR: --ten-bit-input requires support for OpenGL compute shaders\n");
 480                         fprintf(stderr, "       (OpenGL 4.3, or GL_ARB_compute_shader + GL_ARB_shader_image_load_store).\n");
 481                         abort();
 482                 }
 483                 v210_converter.reset(new v210Converter());
 484
 485                 // These are all the widths listed in the Blackmagic SDK documentation
 486                 // (section 2.7.3, “Display Modes”).
 487                 v210_converter->precompile_shader(720);
 488                 v210_converter->precompile_shader(1280);
 489                 v210_converter->precompile_shader(1920);
 490                 v210_converter->precompile_shader(2048);
 491                 v210_converter->precompile_shader(3840);
 492                 v210_converter->precompile_shader(4096);
 493         }
 494         if (global_flags.ten_bit_output) {
 495                 if (!v210Converter::has_hardware_support()) {
 496                         fprintf(stderr, "ERROR: --ten-bit-output requires support for OpenGL compute shaders\n");
 497                         fprintf(stderr, "       (OpenGL 4.3, or GL_ARB_compute_shader + GL_ARB_shader_image_load_store).\n");
 498                         abort();
 499                 }
 500         }
 501
 502         timecode_renderer.reset(new TimecodeRenderer(resource_pool.get(), global_flags.width, global_flags.height));
 503         display_timecode_in_stream = global_flags.display_timecode_in_stream;
 504         display_timecode_on_stdout = global_flags.display_timecode_on_stdout;
 505
 506         if (global_flags.enable_alsa_output) {
 507                 alsa.reset(new ALSAOutput(OUTPUT_FREQUENCY, /*num_channels=*/2));
 508         }
 509         if (global_flags.output_card != -1) {
 510                 desired_output_card_index = global_flags.output_card;
 511                 set_output_card_internal(global_flags.output_card);
 512         }
 513
 514         output_jitter_history.register_metrics({{ "card", "output" }});
 515
 516         ImageInput::start_update_thread(image_update_surface);
 517 }
 518
 519 Mixer::~Mixer()
 520 {
 521         ImageInput::end_update_thread();
 522
 523         if (mjpeg_encoder != nullptr) {
 524                 mjpeg_encoder->stop();
 525         }
 526         httpd.stop();
 527         BMUSBCapture::stop_bm_thread();
 528
 529         for (unsigned card_index = 0; card_index < num_cards + num_video_inputs + num_html_inputs; ++card_index) {
 530                 cards[card_index].capture->stop_dequeue_thread();
 531                 if (cards[card_index].output) {
 532                         cards[card_index].output->end_output();
 533                         cards[card_index].output.reset();
 534                 }
 535         }
 536
 537         video_encoder.reset(nullptr);
 538 }
 539
 540 void Mixer::configure_card(unsigned card_index, CaptureInterface *capture, CardType card_type, DeckLinkOutput *output)
 541 {
 542         printf("Configuring card %d...\n", card_index);
 543
 544         CaptureCard *card = &cards[card_index];
 545         if (card->capture != nullptr) {
 546                 card->capture->stop_dequeue_thread();
 547         }
 548         card->capture.reset(capture);
 549         card->is_fake_capture = (card_type == CardType::FAKE_CAPTURE);
 550         card->is_cef_capture = (card_type == CardType::CEF_INPUT);
 551         card->may_have_dropped_last_frame = false;
 552         card->type = card_type;
 553         if (card->output.get() != output) {
 554                 card->output.reset(output);
 555         }
 556
 557         PixelFormat pixel_format;
 558         if (card_type == CardType::FFMPEG_INPUT) {
 559                 pixel_format = capture->get_current_pixel_format();
 560         } else if (card_type == CardType::CEF_INPUT) {
 561                 pixel_format = PixelFormat_8BitBGRA;
 562         } else if (global_flags.ten_bit_input) {
 563                 pixel_format = PixelFormat_10BitYCbCr;
 564         } else {
 565                 pixel_format = PixelFormat_8BitYCbCr;
 566         }
 567
 568         card->capture->set_frame_callback(bind(&Mixer::bm_frame, this, card_index, _1, _2, _3, _4, _5, _6, _7));
 569         if (card->frame_allocator == nullptr) {
 570                 card->frame_allocator.reset(new PBOFrameAllocator(pixel_format, 8 << 20, global_flags.width, global_flags.height, card_index, mjpeg_encoder.get()));  // 8 MB.
 571         }
 572         card->capture->set_video_frame_allocator(card->frame_allocator.get());
 573         if (card->surface == nullptr) {
 574                 card->surface = create_surface_with_same_format(mixer_surface);
 575         }
 576         while (!card->new_frames.empty()) card->new_frames.pop_front();
 577         card->last_timecode = -1;
 578         card->capture->set_pixel_format(pixel_format);
 579         card->capture->configure_card();
 580
 581         // NOTE: start_bm_capture() happens in thread_func().
 582
 583         DeviceSpec device;
 584         if (card_type == CardType::FFMPEG_INPUT) {
 585                 device = DeviceSpec{InputSourceType::FFMPEG_VIDEO_INPUT, card_index - num_cards};
 586         } else {
 587                 device = DeviceSpec{InputSourceType::CAPTURE_CARD, card_index};
 588         }
 589         audio_mixer->reset_resampler(device);
 590         audio_mixer->set_display_name(device, card->capture->get_description());
 591         audio_mixer->trigger_state_changed_callback();
 592
 593         // Unregister old metrics, if any.
 594         if (!card->labels.empty()) {
 595                 const vector<pair<string, string>> &labels = card->labels;
 596                 card->jitter_history.unregister_metrics(labels);
 597                 card->queue_length_policy.unregister_metrics(labels);
 598                 global_metrics.remove("input_received_frames", labels);
 599                 global_metrics.remove("input_dropped_frames_jitter", labels);
 600                 global_metrics.remove("input_dropped_frames_error", labels);
 601                 global_metrics.remove("input_dropped_frames_resets", labels);
 602                 global_metrics.remove("input_queue_length_frames", labels);
 603                 global_metrics.remove("input_queue_duped_frames", labels);
 604
 605                 global_metrics.remove("input_has_signal_bool", labels);
 606                 global_metrics.remove("input_is_connected_bool", labels);
 607                 global_metrics.remove("input_interlaced_bool", labels);
 608                 global_metrics.remove("input_width_pixels", labels);
 609                 global_metrics.remove("input_height_pixels", labels);
 610                 global_metrics.remove("input_frame_rate_nom", labels);
 611                 global_metrics.remove("input_frame_rate_den", labels);
 612                 global_metrics.remove("input_sample_rate_hz", labels);
 613         }
 614
 615         // Register metrics.
 616         vector<pair<string, string>> labels;
 617         char card_name[64];
 618         snprintf(card_name, sizeof(card_name), "%d", card_index);
 619         labels.emplace_back("card", card_name);
 620
 621         switch (card_type) {
 622         case CardType::LIVE_CARD:
 623                 labels.emplace_back("cardtype", "live");
 624                 break;
 625         case CardType::FAKE_CAPTURE:
 626                 labels.emplace_back("cardtype", "fake");
 627                 break;
 628         case CardType::FFMPEG_INPUT:
 629                 labels.emplace_back("cardtype", "ffmpeg");
 630                 break;
 631         case CardType::CEF_INPUT:
 632                 labels.emplace_back("cardtype", "cef");
 633                 break;
 634         default:
 635                 assert(false);
 636         }
 637         card->jitter_history.register_metrics(labels);
 638         card->queue_length_policy.register_metrics(labels);
 639         global_metrics.add("input_received_frames", labels, &card->metric_input_received_frames);
 640         global_metrics.add("input_dropped_frames_jitter", labels, &card->metric_input_dropped_frames_jitter);
 641         global_metrics.add("input_dropped_frames_error", labels, &card->metric_input_dropped_frames_error);
 642         global_metrics.add("input_dropped_frames_resets", labels, &card->metric_input_resets);
 643         global_metrics.add("input_queue_length_frames", labels, &card->metric_input_queue_length_frames, Metrics::TYPE_GAUGE);
 644         global_metrics.add("input_queue_duped_frames", labels, &card->metric_input_duped_frames);
 645
 646         global_metrics.add("input_has_signal_bool", labels, &card->metric_input_has_signal_bool, Metrics::TYPE_GAUGE);
 647         global_metrics.add("input_is_connected_bool", labels, &card->metric_input_is_connected_bool, Metrics::TYPE_GAUGE);
 648         global_metrics.add("input_interlaced_bool", labels, &card->metric_input_interlaced_bool, Metrics::TYPE_GAUGE);
 649         global_metrics.add("input_width_pixels", labels, &card->metric_input_width_pixels, Metrics::TYPE_GAUGE);
 650         global_metrics.add("input_height_pixels", labels, &card->metric_input_height_pixels, Metrics::TYPE_GAUGE);
 651         global_metrics.add("input_frame_rate_nom", labels, &card->metric_input_frame_rate_nom, Metrics::TYPE_GAUGE);
 652         global_metrics.add("input_frame_rate_den", labels, &card->metric_input_frame_rate_den, Metrics::TYPE_GAUGE);
 653         global_metrics.add("input_sample_rate_hz", labels, &card->metric_input_sample_rate_hz, Metrics::TYPE_GAUGE);
 654         card->labels = labels;
 655 }
 656
 657 void Mixer::set_output_card_internal(int card_index)
 658 {
 659         // We don't really need to take card_mutex, since we're in the mixer
 660         // thread and don't mess with any queues (which is the only thing that happens
 661         // from other threads), but it's probably the safest in the long run.
 662         unique_lock<mutex> lock(card_mutex);
 663         if (output_card_index != -1) {
 664                 // Switch the old card from output to input.
 665                 CaptureCard *old_card = &cards[output_card_index];
 666                 old_card->output->end_output();
 667
 668                 // Stop the fake card that we put into place.
 669                 // This needs to _not_ happen under the mutex, to avoid deadlock
 670                 // (delivering the last frame needs to take the mutex).
 671                 CaptureInterface *fake_capture = old_card->capture.get();
 672                 lock.unlock();
 673                 fake_capture->stop_dequeue_thread();
 674                 lock.lock();
 675                 old_card->capture = move(old_card->parked_capture);  // TODO: reset the metrics
 676                 old_card->is_fake_capture = false;
 677                 old_card->capture->start_bm_capture();
 678         }
 679         if (card_index != -1) {
 680                 CaptureCard *card = &cards[card_index];
 681                 CaptureInterface *capture = card->capture.get();
 682                 // TODO: DeckLinkCapture::stop_dequeue_thread can actually take
 683                 // several seconds to complete (blocking on DisableVideoInput);
 684                 // see if we can maybe do it asynchronously.
 685                 lock.unlock();
 686                 capture->stop_dequeue_thread();
 687                 lock.lock();
 688                 card->parked_capture = move(card->capture);
 689                 CaptureInterface *fake_capture = new FakeCapture(global_flags.width, global_flags.height, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
 690                 configure_card(card_index, fake_capture, CardType::FAKE_CAPTURE, card->output.release());
 691                 card->queue_length_policy.reset(card_index);
 692                 card->capture->start_bm_capture();
 693                 desired_output_video_mode = output_video_mode = card->output->pick_video_mode(desired_output_video_mode);
 694                 card->output->start_output(desired_output_video_mode, pts_int);
 695         }
 696         output_card_index = card_index;
 697         output_jitter_history.clear();
 698 }
 699
 700 namespace {
 701
 702 int unwrap_timecode(uint16_t current_wrapped, int last)
 703 {
 704         uint16_t last_wrapped = last & 0xffff;
 705         if (current_wrapped > last_wrapped) {
 706                 return (last & ~0xffff) | current_wrapped;
 707         } else {
 708                 return 0x10000 + ((last & ~0xffff) | current_wrapped);
 709         }
 710 }
 711
 712 DeviceSpec card_index_to_device(unsigned card_index, unsigned num_cards)
 713 {
 714         if (card_index >= num_cards) {
 715                 return DeviceSpec{InputSourceType::FFMPEG_VIDEO_INPUT, card_index - num_cards};
 716         } else {
 717                 return DeviceSpec{InputSourceType::CAPTURE_CARD, card_index};
 718         }
 719 }
 720
 721 }  // namespace
 722
 723 void Mixer::bm_frame(unsigned card_index, uint16_t timecode,
 724                      FrameAllocator::Frame video_frame, size_t video_offset, VideoFormat video_format,
 725                      FrameAllocator::Frame audio_frame, size_t audio_offset, AudioFormat audio_format)
 726 {
 727         DeviceSpec device = card_index_to_device(card_index, num_cards);
 728         CaptureCard *card = &cards[card_index];
 729
 730         ++card->metric_input_received_frames;
 731         card->metric_input_has_signal_bool = video_format.has_signal;
 732         card->metric_input_is_connected_bool = video_format.is_connected;
 733         card->metric_input_interlaced_bool = video_format.interlaced;
 734         card->metric_input_width_pixels = video_format.width;
 735         card->metric_input_height_pixels = video_format.height;
 736         card->metric_input_frame_rate_nom = video_format.frame_rate_nom;
 737         card->metric_input_frame_rate_den = video_format.frame_rate_den;
 738         card->metric_input_sample_rate_hz = audio_format.sample_rate;
 739
 740         if (is_mode_scanning[card_index]) {
 741                 if (video_format.has_signal) {
 742                         // Found a stable signal, so stop scanning.
 743                         is_mode_scanning[card_index] = false;
 744                 } else {
 745                         static constexpr double switch_time_s = 0.1;  // Should be enough time for the signal to stabilize.
 746                         steady_clock::time_point now = steady_clock::now();
 747                         double sec_since_last_switch = duration<double>(steady_clock::now() - last_mode_scan_change[card_index]).count();
 748                         if (sec_since_last_switch > switch_time_s) {
 749                                 // It isn't this mode; try the next one.
 750                                 mode_scanlist_index[card_index]++;
 751                                 mode_scanlist_index[card_index] %= mode_scanlist[card_index].size();
 752                                 cards[card_index].capture->set_video_mode(mode_scanlist[card_index][mode_scanlist_index[card_index]]);
 753                                 last_mode_scan_change[card_index] = now;
 754                         }
 755                 }
 756         }
 757
 758         int64_t frame_length = int64_t(TIMEBASE) * video_format.frame_rate_den / video_format.frame_rate_nom;
 759         assert(frame_length > 0);
 760
 761         size_t num_samples = (audio_frame.len > audio_offset) ? (audio_frame.len - audio_offset) / audio_format.num_channels / (audio_format.bits_per_sample / 8) : 0;
 762         if (num_samples > OUTPUT_FREQUENCY / 10 && card->type != CardType::FFMPEG_INPUT) {
 763                 printf("%s: Dropping frame with implausible audio length (len=%d, offset=%d) [timecode=0x%04x video_len=%d video_offset=%d video_format=%x)\n",
 764                         spec_to_string(device).c_str(), int(audio_frame.len), int(audio_offset),
 765                         timecode, int(video_frame.len), int(video_offset), video_format.id);
 766                 if (video_frame.owner) {
 767                         video_frame.owner->release_frame(video_frame);
 768                 }
 769                 if (audio_frame.owner) {
 770                         audio_frame.owner->release_frame(audio_frame);
 771                 }
 772                 return;
 773         }
 774
 775         int dropped_frames = 0;
 776         if (card->last_timecode != -1) {
 777                 dropped_frames = unwrap_timecode(timecode, card->last_timecode) - card->last_timecode - 1;
 778         }
 779
 780         // Number of samples per frame if we need to insert silence.
 781         // (Could be nonintegral, but resampling will save us then.)
 782         const int silence_samples = OUTPUT_FREQUENCY * video_format.frame_rate_den / video_format.frame_rate_nom;
 783
 784         if (dropped_frames > MAX_FPS * 2) {
 785                 fprintf(stderr, "%s lost more than two seconds (or time code jumping around; from 0x%04x to 0x%04x), resetting resampler\n",
 786                         spec_to_string(device).c_str(), card->last_timecode, timecode);
 787                 audio_mixer->reset_resampler(device);
 788                 dropped_frames = 0;
 789                 ++card->metric_input_resets;
 790         } else if (dropped_frames > 0) {
 791                 // Insert silence as needed.
 792                 fprintf(stderr, "%s dropped %d frame(s) (before timecode 0x%04x), inserting silence.\n",
 793                         spec_to_string(device).c_str(), dropped_frames, timecode);
 794                 card->metric_input_dropped_frames_error += dropped_frames;
 795
 796                 bool success;
 797                 do {
 798                         success = audio_mixer->add_silence(device, silence_samples, dropped_frames);
 799                 } while (!success);
 800         }
 801
 802         if (num_samples > 0) {
 803                 audio_mixer->add_audio(device, audio_frame.data + audio_offset, num_samples, audio_format, audio_frame.received_timestamp);
 804
 805                 // Audio for the MJPEG stream. We don't resample; audio that's not in 48 kHz
 806                 // just gets dropped for now.
 807                 //
 808                 // Only bother doing MJPEG encoding if there are any connected clients
 809                 // that want the stream.
 810                 if (httpd.get_num_connected_multicam_clients() > 0 ||
 811                     httpd.get_num_connected_siphon_clients(card_index) > 0) {
 812                         vector<int32_t> converted_samples = convert_audio_to_fixed32(audio_frame.data + audio_offset, num_samples, audio_format, 2);
 813                         lock_guard<mutex> lock(card_mutex);
 814                         if (card->new_raw_audio.empty()) {
 815                                 card->new_raw_audio = move(converted_samples);
 816                         } else {
 817                                 // For raw audio, we don't really synchronize audio and video;
 818                                 // we just put the audio in frame by frame, and if a video frame is
 819                                 // dropped, we still keep the audio, which means it will be added
 820                                 // to the beginning of the next frame. It would probably be better
 821                                 // to move the audio pts earlier to show this, but most players can
 822                                 // live with some jitter, and in a lot of ways, it's much nicer for
 823                                 // Futatabi to have all audio locked to a video frame.
 824                                 card->new_raw_audio.insert(card->new_raw_audio.end(), converted_samples.begin(), converted_samples.end());
 825
 826                                 // Truncate to one second, just to be sure we don't have infinite buildup in case of weirdness.
 827                                 if (card->new_raw_audio.size() > OUTPUT_FREQUENCY * 2) {
 828                                         size_t excess_samples = card->new_raw_audio.size() - OUTPUT_FREQUENCY * 2;
 829                                         card->new_raw_audio.erase(card->new_raw_audio.begin(), card->new_raw_audio.begin() + excess_samples);
 830                                 }
 831                         }
 832                 }
 833         }
 834
 835         // Done with the audio, so release it.
 836         if (audio_frame.owner) {
 837                 audio_frame.owner->release_frame(audio_frame);
 838         }
 839
 840         card->last_timecode = timecode;
 841
 842         PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)video_frame.userdata;
 843         if (card->type == CardType::FFMPEG_INPUT && userdata != nullptr) {
 844                 FFmpegCapture *ffmpeg_capture = static_cast<FFmpegCapture *>(card->capture.get());
 845                 userdata->has_last_subtitle = ffmpeg_capture->get_has_last_subtitle();
 846                 userdata->last_subtitle = ffmpeg_capture->get_last_subtitle();
 847         }
 848
 849         size_t cbcr_width, cbcr_height, cbcr_offset, y_offset;
 850         size_t expected_length = video_format.stride * (video_format.height + video_format.extra_lines_top + video_format.extra_lines_bottom);
 851         if (userdata != nullptr && userdata->pixel_format == PixelFormat_8BitYCbCrPlanar) {
 852                 // The calculation above is wrong for planar Y'CbCr, so just override it.
 853                 assert(card->type == CardType::FFMPEG_INPUT);
 854                 assert(video_offset == 0);
 855                 expected_length = video_frame.len;
 856
 857                 userdata->ycbcr_format = (static_cast<FFmpegCapture *>(card->capture.get()))->get_current_frame_ycbcr_format();
 858                 cbcr_width = video_format.width / userdata->ycbcr_format.chroma_subsampling_x;
 859                 cbcr_height = video_format.height / userdata->ycbcr_format.chroma_subsampling_y;
 860                 cbcr_offset = video_format.width * video_format.height;
 861                 y_offset = 0;
 862         } else {
 863                 // All the other Y'CbCr formats are 4:2:2.
 864                 cbcr_width = video_format.width / 2;
 865                 cbcr_height = video_format.height;
 866                 cbcr_offset = video_offset / 2;
 867                 y_offset = video_frame.size / 2 + video_offset / 2;
 868         }
 869         if (video_frame.len - video_offset == 0 ||
 870             video_frame.len - video_offset != expected_length) {
 871                 if (video_frame.len != 0) {
 872                         printf("%s: Dropping video frame with wrong length (%zu; expected %zu)\n",
 873                                 spec_to_string(device).c_str(), video_frame.len - video_offset, expected_length);
 874                 }
 875                 if (video_frame.owner) {
 876                         video_frame.owner->release_frame(video_frame);
 877                 }
 878
 879                 // Still send on the information that we _had_ a frame, even though it's corrupted,
 880                 // so that pts can go up accordingly.
 881                 {
 882                         lock_guard<mutex> lock(card_mutex);
 883                         CaptureCard::NewFrame new_frame;
 884                         new_frame.frame = RefCountedFrame(FrameAllocator::Frame());
 885                         new_frame.length = frame_length;
 886                         new_frame.interlaced = false;
 887                         new_frame.dropped_frames = dropped_frames;
 888                         new_frame.received_timestamp = video_frame.received_timestamp;
 889                         card->new_frames.push_back(move(new_frame));
 890                         card->jitter_history.frame_arrived(video_frame.received_timestamp, frame_length, dropped_frames);
 891                 }
 892                 card->new_frames_changed.notify_all();
 893                 return;
 894         }
 895
 896         unsigned num_fields = video_format.interlaced ? 2 : 1;
 897         steady_clock::time_point frame_upload_start;
 898         bool interlaced_stride = false;
 899         if (video_format.interlaced) {
 900                 // Send the two fields along as separate frames; the other side will need to add
 901                 // a deinterlacer to actually get this right.
 902                 assert(video_format.height % 2 == 0);
 903                 video_format.height /= 2;
 904                 cbcr_height /= 2;
 905                 assert(frame_length % 2 == 0);
 906                 frame_length /= 2;
 907                 num_fields = 2;
 908                 if (video_format.second_field_start == 1) {
 909                         interlaced_stride = true;
 910                 }
 911                 frame_upload_start = steady_clock::now();
 912         }
 913         assert(userdata != nullptr);
 914         userdata->last_interlaced = video_format.interlaced;
 915         userdata->last_has_signal = video_format.has_signal;
 916         userdata->last_is_connected = video_format.is_connected;
 917         userdata->last_frame_rate_nom = video_format.frame_rate_nom;
 918         userdata->last_frame_rate_den = video_format.frame_rate_den;
 919         RefCountedFrame frame(video_frame);
 920
 921         // Upload the textures.
 922         for (unsigned field = 0; field < num_fields; ++field) {
 923                 // Put the actual texture upload in a lambda that is executed in the main thread.
 924                 // It is entirely possible to do this in the same thread (and it might even be
 925                 // faster, depending on the GPU and driver), but it appears to be trickling
 926                 // driver bugs very easily.
 927                 //
 928                 // Note that this means we must hold on to the actual frame data in <userdata>
 929                 // until the upload command is run, but we hold on to <frame> much longer than that
 930                 // (in fact, all the way until we no longer use the texture in rendering).
 931                 auto upload_func = [this, field, video_format, y_offset, video_offset, cbcr_offset, cbcr_width, cbcr_height, interlaced_stride, userdata]() {
 932                         unsigned field_start_line;
 933                         if (field == 1) {
 934                                 field_start_line = video_format.second_field_start;
 935                         } else {
 936                                 field_start_line = video_format.extra_lines_top;
 937                         }
 938
 939                         // For anything not FRAME_FORMAT_YCBCR_10BIT, v210_width will be nonsensical but not used.
 940                         size_t v210_width = video_format.stride / sizeof(uint32_t);
 941                         ensure_texture_resolution(userdata, field, video_format.width, video_format.height, cbcr_width, cbcr_height, v210_width);
 942
 943                         glBindBuffer(GL_PIXEL_UNPACK_BUFFER, userdata->pbo);
 944                         check_error();
 945
 946                         switch (userdata->pixel_format) {
 947                         case PixelFormat_10BitYCbCr: {
 948                                 size_t field_start = video_offset + video_format.stride * field_start_line;
 949                                 upload_texture(userdata->tex_v210[field], v210_width, video_format.height, video_format.stride, interlaced_stride, GL_RGBA, GL_UNSIGNED_INT_2_10_10_10_REV, field_start);
 950                                 v210_converter->convert(userdata->tex_v210[field], userdata->tex_444[field], video_format.width, video_format.height);
 951                                 break;
 952                         }
 953                         case PixelFormat_8BitYCbCr: {
 954                                 size_t field_y_start = y_offset + video_format.width * field_start_line;
 955                                 size_t field_cbcr_start = cbcr_offset + cbcr_width * field_start_line * sizeof(uint16_t);
 956
 957                                 // Make up our own strides, since we are interleaving.
 958                                 upload_texture(userdata->tex_y[field], video_format.width, video_format.height, video_format.width, interlaced_stride, GL_RED, GL_UNSIGNED_BYTE, field_y_start);
 959                                 upload_texture(userdata->tex_cbcr[field], cbcr_width, cbcr_height, cbcr_width * sizeof(uint16_t), interlaced_stride, GL_RG, GL_UNSIGNED_BYTE, field_cbcr_start);
 960                                 break;
 961                         }
 962                         case PixelFormat_8BitYCbCrPlanar: {
 963                                 assert(field_start_line == 0);  // We don't really support interlaced here.
 964                                 size_t field_y_start = y_offset;
 965                                 size_t field_cb_start = cbcr_offset;
 966                                 size_t field_cr_start = cbcr_offset + cbcr_width * cbcr_height;
 967
 968                                 // Make up our own strides, since we are interleaving.
 969                                 upload_texture(userdata->tex_y[field], video_format.width, video_format.height, video_format.width, interlaced_stride, GL_RED, GL_UNSIGNED_BYTE, field_y_start);
 970                                 upload_texture(userdata->tex_cb[field], cbcr_width, cbcr_height, cbcr_width, interlaced_stride, GL_RED, GL_UNSIGNED_BYTE, field_cb_start);
 971                                 upload_texture(userdata->tex_cr[field], cbcr_width, cbcr_height, cbcr_width, interlaced_stride, GL_RED, GL_UNSIGNED_BYTE, field_cr_start);
 972                                 break;
 973                         }
 974                         case PixelFormat_8BitBGRA: {
 975                                 size_t field_start = video_offset + video_format.stride * field_start_line;
 976                                 upload_texture(userdata->tex_rgba[field], video_format.width, video_format.height, video_format.stride, interlaced_stride, GL_BGRA, GL_UNSIGNED_INT_8_8_8_8_REV, field_start);
 977                                 // These could be asked to deliver mipmaps at any time.
 978                                 glBindTexture(GL_TEXTURE_2D, userdata->tex_rgba[field]);
 979                                 check_error();
 980                                 glGenerateMipmap(GL_TEXTURE_2D);
 981                                 check_error();
 982                                 glBindTexture(GL_TEXTURE_2D, 0);
 983                                 check_error();
 984                                 break;
 985                         }
 986                         default:
 987                                 assert(false);
 988                         }
 989
 990                         glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
 991                         check_error();
 992                 };
 993
 994                 if (field == 1) {
 995                         // Don't upload the second field as fast as we can; wait until
 996                         // the field time has approximately passed. (Otherwise, we could
 997                         // get timing jitter against the other sources, and possibly also
 998                         // against the video display, although the latter is not as critical.)
 999                         // This requires our system clock to be reasonably close to the
1000                         // video clock, but that's not an unreasonable assumption.
1001                         steady_clock::time_point second_field_start = frame_upload_start +
1002                                 nanoseconds(frame_length * 1000000000 / TIMEBASE);
1003                         this_thread::sleep_until(second_field_start);
1004                 }
1005
1006                 {
1007                         lock_guard<mutex> lock(card_mutex);
1008                         CaptureCard::NewFrame new_frame;
1009                         new_frame.frame = frame;
1010                         new_frame.length = frame_length;
1011                         new_frame.field = field;
1012                         new_frame.interlaced = video_format.interlaced;
1013                         new_frame.upload_func = upload_func;
1014                         new_frame.dropped_frames = dropped_frames;
1015                         new_frame.received_timestamp = video_frame.received_timestamp;  // Ignore the audio timestamp.
1016                         new_frame.video_format = video_format;
1017                         new_frame.y_offset = y_offset;
1018                         new_frame.cbcr_offset = cbcr_offset;
1019                         if (card->type == CardType::FFMPEG_INPUT) {
1020                                 FFmpegCapture *ffmpeg_capture = static_cast<FFmpegCapture *>(card->capture.get());
1021                                 new_frame.neutral_color = ffmpeg_capture->get_last_neutral_color();
1022                         }
1023                         card->new_frames.push_back(move(new_frame));
1024                         card->jitter_history.frame_arrived(video_frame.received_timestamp, frame_length, dropped_frames);
1025                         card->may_have_dropped_last_frame = false;
1026                 }
1027                 card->new_frames_changed.notify_all();
1028         }
1029 }
1030
1031 void Mixer::bm_hotplug_add(libusb_device *dev)
1032 {
1033         lock_guard<mutex> lock(hotplug_mutex);
1034         hotplugged_cards.push_back(dev);
1035 }
1036
1037 void Mixer::bm_hotplug_remove(unsigned card_index)
1038 {
1039         cards[card_index].new_frames_changed.notify_all();
1040 }
1041
1042 void Mixer::thread_func()
1043 {
1044         pthread_setname_np(pthread_self(), "Mixer_OpenGL");
1045
1046         eglBindAPI(EGL_OPENGL_API);
1047         QOpenGLContext *context = create_context(mixer_surface);
1048         if (!make_current(context, mixer_surface)) {
1049                 printf("oops\n");
1050                 abort();
1051         }
1052
1053         // Start the actual capture. (We don't want to do it before we're actually ready
1054         // to process output frames.)
1055         for (unsigned card_index = 0; card_index < num_cards + num_video_inputs + num_html_inputs; ++card_index) {
1056                 if (int(card_index) != output_card_index) {
1057                         cards[card_index].capture->start_bm_capture();
1058                 }
1059         }
1060
1061         BasicStats basic_stats(/*verbose=*/true, /*use_opengl=*/true);
1062         int stats_dropped_frames = 0;
1063
1064         while (!should_quit) {
1065                 if (desired_output_card_index != output_card_index) {
1066                         set_output_card_internal(desired_output_card_index);
1067                 }
1068                 if (output_card_index != -1 &&
1069                     desired_output_video_mode != output_video_mode) {
1070                         DeckLinkOutput *output = cards[output_card_index].output.get();
1071                         output->end_output();
1072                         desired_output_video_mode = output_video_mode = output->pick_video_mode(desired_output_video_mode);
1073                         output->start_output(desired_output_video_mode, pts_int);
1074                 }
1075
1076                 CaptureCard::NewFrame new_frames[MAX_VIDEO_CARDS];
1077                 bool has_new_frame[MAX_VIDEO_CARDS] = { false };
1078
1079                 bool master_card_is_output;
1080                 unsigned master_card_index;
1081                 if (output_card_index != -1) {
1082                         master_card_is_output = true;
1083                         master_card_index = output_card_index;
1084                 } else {
1085                         master_card_is_output = false;
1086                         master_card_index = theme->map_signal_to_card(master_clock_channel);
1087                         assert(master_card_index < num_cards + num_video_inputs);
1088                 }
1089
1090                 vector<int32_t> raw_audio[MAX_VIDEO_CARDS];  // For MJPEG encoding.
1091                 OutputFrameInfo output_frame_info = get_one_frame_from_each_card(master_card_index, master_card_is_output, new_frames, has_new_frame, raw_audio);
1092                 schedule_audio_resampling_tasks(output_frame_info.dropped_frames, output_frame_info.num_samples, output_frame_info.frame_duration, output_frame_info.is_preroll, output_frame_info.frame_timestamp);
1093                 stats_dropped_frames += output_frame_info.dropped_frames;
1094
1095                 handle_hotplugged_cards();
1096
1097                 for (unsigned card_index = 0; card_index < num_cards + num_video_inputs + num_html_inputs; ++card_index) {
1098                         DeviceSpec device = card_index_to_device(card_index, num_cards);
1099                         if (card_index == master_card_index || !has_new_frame[card_index]) {
1100                                 continue;
1101                         }
1102                         if (new_frames[card_index].frame->len == 0) {
1103                                 ++new_frames[card_index].dropped_frames;
1104                         }
1105                         if (new_frames[card_index].dropped_frames > 0) {
1106                                 printf("%s dropped %d frames before this\n",
1107                                         spec_to_string(device).c_str(), int(new_frames[card_index].dropped_frames));
1108                         }
1109                 }
1110
1111                 // If the first card is reporting a corrupted or otherwise dropped frame,
1112                 // just increase the pts (skipping over this frame) and don't try to compute anything new.
1113                 if (!master_card_is_output && new_frames[master_card_index].frame->len == 0) {
1114                         ++stats_dropped_frames;
1115                         pts_int += new_frames[master_card_index].length;
1116                         continue;
1117                 }
1118
1119                 for (unsigned card_index = 0; card_index < num_cards + num_video_inputs + num_html_inputs; ++card_index) {
1120                         if (!has_new_frame[card_index] || new_frames[card_index].frame->len == 0)
1121                                 continue;
1122
1123                         CaptureCard::NewFrame *new_frame = &new_frames[card_index];
1124                         assert(new_frame->frame != nullptr);
1125                         insert_new_frame(new_frame->frame, new_frame->field, new_frame->interlaced, card_index, &input_state);
1126                         check_error();
1127
1128                         // The new texture might need uploading before use.
1129                         if (new_frame->upload_func) {
1130                                 new_frame->upload_func();
1131                                 new_frame->upload_func = nullptr;
1132                         }
1133
1134                         // Only set the white balance if it actually changed. This means that the user
1135                         // is free to override the white balance in a video with no white balance information
1136                         // actually set (ie. r=g=b=1 all the time), or one where the white point is wrong,
1137                         // but frame-to-frame decisions will be heeded. We do this pretty much as late
1138                         // as possible (ie., after picking out the frame from the buffer), so that we are sure
1139                         // that the change takes effect on exactly the right frame.
1140                         if (fabs(new_frame->neutral_color.r - last_received_neutral_color[card_index].r) > 1e-3 ||
1141                             fabs(new_frame->neutral_color.g - last_received_neutral_color[card_index].g) > 1e-3 ||
1142                             fabs(new_frame->neutral_color.b - last_received_neutral_color[card_index].b) > 1e-3) {
1143                                 theme->set_wb_for_card(card_index, new_frame->neutral_color.r, new_frame->neutral_color.g, new_frame->neutral_color.b);
1144                                 last_received_neutral_color[card_index] = new_frame->neutral_color;
1145                         }
1146
1147                         if (new_frame->frame->data_copy != nullptr && mjpeg_encoder->should_encode_mjpeg_for_card(card_index)) {
1148                                 RGBTriplet neutral_color = theme->get_white_balance_for_card(card_index);
1149                                 mjpeg_encoder->upload_frame(pts_int, card_index, new_frame->frame, new_frame->video_format, new_frame->y_offset, new_frame->cbcr_offset, move(raw_audio[card_index]), neutral_color);
1150                         }
1151
1152                 }
1153
1154                 int64_t frame_duration = output_frame_info.frame_duration;
1155                 render_one_frame(frame_duration);
1156                 {
1157                         lock_guard<mutex> lock(frame_num_mutex);
1158                         ++frame_num;
1159                 }
1160                 frame_num_updated.notify_all();
1161                 pts_int += frame_duration;
1162
1163                 basic_stats.update(frame_num, stats_dropped_frames);
1164                 // if (frame_num % 100 == 0) chain->print_phase_timing();
1165
1166                 if (should_cut.exchange(false)) {  // Test and clear.
1167                         video_encoder->do_cut(frame_num);
1168                 }
1169
1170 #if 0
1171                 // Reset every 100 frames, so that local variations in frame times
1172                 // (especially for the first few frames, when the shaders are
1173                 // compiled etc.) don't make it hard to measure for the entire
1174                 // remaining duration of the program.
1175                 if (frame == 10000) {
1176                         frame = 0;
1177                         start = now;
1178                 }
1179 #endif
1180                 check_error();
1181         }
1182
1183         resource_pool->clean_context();
1184 }
1185
1186 bool Mixer::input_card_is_master_clock(unsigned card_index, unsigned master_card_index) const
1187 {
1188         if (output_card_index != -1) {
1189                 // The output card (ie., cards[output_card_index].output) is the master clock,
1190                 // so no input card (ie., cards[card_index].capture) is.
1191                 return false;
1192         }
1193         return (card_index == master_card_index);
1194 }
1195
1196 void Mixer::trim_queue(CaptureCard *card, size_t safe_queue_length)
1197 {
1198         // Count the number of frames in the queue, including any frames
1199         // we dropped. It's hard to know exactly how we should deal with
1200         // dropped (corrupted) input frames; they don't help our goal of
1201         // avoiding starvation, but they still add to the problem of latency.
1202         // Since dropped frames is going to mean a bump in the signal anyway,
1203         // we err on the side of having more stable latency instead.
1204         unsigned queue_length = 0;
1205         for (const CaptureCard::NewFrame &frame : card->new_frames) {
1206                 queue_length += frame.dropped_frames + 1;
1207         }
1208
1209         // If needed, drop frames until the queue is below the safe limit.
1210         // We prefer to drop from the head, because all else being equal,
1211         // we'd like more recent frames (less latency).
1212         unsigned dropped_frames = 0;
1213         while (queue_length > safe_queue_length) {
1214                 assert(!card->new_frames.empty());
1215                 assert(queue_length > card->new_frames.front().dropped_frames);
1216                 queue_length -= card->new_frames.front().dropped_frames;
1217
1218                 if (queue_length <= safe_queue_length) {
1219                         // No need to drop anything.
1220                         break;
1221                 }
1222
1223                 card->new_frames.pop_front();
1224                 card->new_frames_changed.notify_all();
1225                 --queue_length;
1226                 ++dropped_frames;
1227
1228                 if (queue_length == 0 && card->is_cef_capture) {
1229                         card->may_have_dropped_last_frame = true;
1230                 }
1231         }
1232
1233         card->metric_input_dropped_frames_jitter += dropped_frames;
1234         card->metric_input_queue_length_frames = queue_length;
1235
1236 #if 0
1237         if (dropped_frames > 0) {
1238                 fprintf(stderr, "Card %u dropped %u frame(s) to keep latency down.\n",
1239                         card_index, dropped_frames);
1240         }
1241 #endif
1242 }
1243
1244 pair<string, string> Mixer::get_channels_json()
1245 {
1246         Channels ret;
1247         for (int channel_idx = 0; channel_idx < theme->get_num_channels(); ++channel_idx) {
1248                 Channel *channel = ret.add_channel();
1249                 channel->set_index(channel_idx + 2);
1250                 channel->set_name(theme->get_channel_name(channel_idx + 2));
1251                 channel->set_color(theme->get_channel_color(channel_idx + 2));
1252         }
1253         string contents;
1254         google::protobuf::util::MessageToJsonString(ret, &contents);  // Ignore any errors.
1255         return make_pair(contents, "text/json");
1256 }
1257
1258 pair<string, string> Mixer::get_channel_color_http(unsigned channel_idx)
1259 {
1260         return make_pair(theme->get_channel_color(channel_idx), "text/plain");
1261 }
1262
1263 Mixer::OutputFrameInfo Mixer::get_one_frame_from_each_card(unsigned master_card_index, bool master_card_is_output, CaptureCard::NewFrame new_frames[MAX_VIDEO_CARDS], bool has_new_frame[MAX_VIDEO_CARDS], vector<int32_t> raw_audio[MAX_VIDEO_CARDS])
1264 {
1265         OutputFrameInfo output_frame_info;
1266 start:
1267         unique_lock<mutex> lock(card_mutex, defer_lock);
1268         if (master_card_is_output) {
1269                 // Clocked to the output, so wait for it to be ready for the next frame.
1270                 cards[master_card_index].output->wait_for_frame(pts_int, &output_frame_info.dropped_frames, &output_frame_info.frame_duration, &output_frame_info.is_preroll, &output_frame_info.frame_timestamp);
1271                 lock.lock();
1272         } else {
1273                 // Wait for the master card to have a new frame.
1274                 // TODO: Add a timeout.
1275                 output_frame_info.is_preroll = false;
1276                 lock.lock();
1277                 cards[master_card_index].new_frames_changed.wait(lock, [this, master_card_index]{ return !cards[master_card_index].new_frames.empty() || cards[master_card_index].capture->get_disconnected(); });
1278         }
1279
1280         if (master_card_is_output) {
1281                 handle_hotplugged_cards();
1282         } else if (cards[master_card_index].new_frames.empty()) {
1283                 // We were woken up, but not due to a new frame. Deal with it
1284                 // and then restart.
1285                 assert(cards[master_card_index].capture->get_disconnected());
1286                 handle_hotplugged_cards();
1287                 lock.unlock();
1288                 goto start;
1289         }
1290
1291         for (unsigned card_index = 0; card_index < num_cards + num_video_inputs + num_html_inputs; ++card_index) {
1292                 CaptureCard *card = &cards[card_index];
1293                 if (card->new_frames.empty()) {  // Starvation.
1294                         ++card->metric_input_duped_frames;
1295 #ifdef HAVE_CEF
1296                         if (card->is_cef_capture && card->may_have_dropped_last_frame) {
1297                                 // Unlike other sources, CEF is not guaranteed to send us a steady
1298                                 // stream of frames, so we'll have to ask it to repaint the frame
1299                                 // we dropped. (may_have_dropped_last_frame is set whenever we
1300                                 // trim the queue completely away, and cleared when we actually
1301                                 // get a new frame.)
1302                                 ((CEFCapture *)card->capture.get())->request_new_frame(/*ignore_if_locked=*/true);
1303                         }
1304 #endif
1305                 } else {
1306                         new_frames[card_index] = move(card->new_frames.front());
1307                         has_new_frame[card_index] = true;
1308                         card->new_frames.pop_front();
1309                         card->new_frames_changed.notify_all();
1310                 }
1311
1312                 raw_audio[card_index] = move(card->new_raw_audio);
1313         }
1314
1315         if (!master_card_is_output) {
1316                 output_frame_info.frame_timestamp = new_frames[master_card_index].received_timestamp;
1317                 output_frame_info.dropped_frames = new_frames[master_card_index].dropped_frames;
1318                 output_frame_info.frame_duration = new_frames[master_card_index].length;
1319         }
1320
1321         if (!output_frame_info.is_preroll) {
1322                 output_jitter_history.frame_arrived(output_frame_info.frame_timestamp, output_frame_info.frame_duration, output_frame_info.dropped_frames);
1323         }
1324
1325         for (unsigned card_index = 0; card_index < num_cards + num_video_inputs + num_html_inputs; ++card_index) {
1326                 CaptureCard *card = &cards[card_index];
1327                 if (has_new_frame[card_index] &&
1328                     !input_card_is_master_clock(card_index, master_card_index) &&
1329                     !output_frame_info.is_preroll) {
1330                         card->queue_length_policy.update_policy(
1331                                 output_frame_info.frame_timestamp,
1332                                 card->jitter_history.get_expected_next_frame(),
1333                                 new_frames[master_card_index].length,
1334                                 output_frame_info.frame_duration,
1335                                 card->jitter_history.estimate_max_jitter(),
1336                                 output_jitter_history.estimate_max_jitter());
1337                         trim_queue(card, min<int>(global_flags.max_input_queue_frames,
1338                                                   card->queue_length_policy.get_safe_queue_length()));
1339                 }
1340         }
1341
1342         // This might get off by a fractional sample when changing master card
1343         // between ones with different frame rates, but that's fine.
1344         int num_samples_times_timebase = OUTPUT_FREQUENCY * output_frame_info.frame_duration + fractional_samples;
1345         output_frame_info.num_samples = num_samples_times_timebase / TIMEBASE;
1346         fractional_samples = num_samples_times_timebase % TIMEBASE;
1347         assert(output_frame_info.num_samples >= 0);
1348
1349         return output_frame_info;
1350 }
1351
1352 void Mixer::handle_hotplugged_cards()
1353 {
1354         // Check for cards that have been disconnected since last frame.
1355         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
1356                 CaptureCard *card = &cards[card_index];
1357                 if (card->capture->get_disconnected()) {
1358                         fprintf(stderr, "Card %u went away, replacing with a fake card.\n", card_index);
1359                         FakeCapture *capture = new FakeCapture(global_flags.width, global_flags.height, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
1360                         configure_card(card_index, capture, CardType::FAKE_CAPTURE, /*output=*/nullptr);
1361                         card->queue_length_policy.reset(card_index);
1362                         card->capture->start_bm_capture();
1363                 }
1364         }
1365
1366         // Check for cards that have been connected since last frame.
1367         vector<libusb_device *> hotplugged_cards_copy;
1368         {
1369                 lock_guard<mutex> lock(hotplug_mutex);
1370                 swap(hotplugged_cards, hotplugged_cards_copy);
1371         }
1372         for (libusb_device *new_dev : hotplugged_cards_copy) {
1373                 // Look for a fake capture card where we can stick this in.
1374                 int free_card_index = -1;
1375                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
1376                         if (cards[card_index].is_fake_capture) {
1377                                 free_card_index = card_index;
1378                                 break;
1379                         }
1380                 }
1381
1382                 if (free_card_index == -1) {
1383                         fprintf(stderr, "New card plugged in, but no free slots -- ignoring.\n");
1384                         libusb_unref_device(new_dev);
1385                 } else {
1386                         // BMUSBCapture takes ownership.
1387                         fprintf(stderr, "New card plugged in, choosing slot %d.\n", free_card_index);
1388                         CaptureCard *card = &cards[free_card_index];
1389                         BMUSBCapture *capture = new BMUSBCapture(free_card_index, new_dev);
1390                         configure_card(free_card_index, capture, CardType::LIVE_CARD, /*output=*/nullptr);
1391                         card->queue_length_policy.reset(free_card_index);
1392                         capture->set_card_disconnected_callback(bind(&Mixer::bm_hotplug_remove, this, free_card_index));
1393                         capture->start_bm_capture();
1394                 }
1395         }
1396 }
1397
1398
1399 void Mixer::schedule_audio_resampling_tasks(unsigned dropped_frames, int num_samples_per_frame, int length_per_frame, bool is_preroll, steady_clock::time_point frame_timestamp)
1400 {
1401         // Resample the audio as needed, including from previously dropped frames.
1402         assert(num_cards > 0);
1403         for (unsigned frame_num = 0; frame_num < dropped_frames + 1; ++frame_num) {
1404                 const bool dropped_frame = (frame_num != dropped_frames);
1405                 {
1406                         // Signal to the audio thread to process this frame.
1407                         // Note that if the frame is a dropped frame, we signal that
1408                         // we don't want to use this frame as base for adjusting
1409                         // the resampler rate. The reason for this is that the timing
1410                         // of these frames is often way too late; they typically don't
1411                         // “arrive” before we synthesize them. Thus, we could end up
1412                         // in a situation where we have inserted e.g. five audio frames
1413                         // into the queue before we then start pulling five of them
1414                         // back out. This makes ResamplingQueue overestimate the delay,
1415                         // causing undue resampler changes. (We _do_ use the last,
1416                         // non-dropped frame; perhaps we should just discard that as well,
1417                         // since dropped frames are expected to be rare, and it might be
1418                         // better to just wait until we have a slightly more normal situation).
1419                         lock_guard<mutex> lock(audio_mutex);
1420                         bool adjust_rate = !dropped_frame && !is_preroll;
1421                         audio_task_queue.push(AudioTask{pts_int, num_samples_per_frame, adjust_rate, frame_timestamp});
1422                         audio_task_queue_changed.notify_one();
1423                 }
1424                 if (dropped_frame) {
1425                         // For dropped frames, increase the pts. Note that if the format changed
1426                         // in the meantime, we have no way of detecting that; we just have to
1427                         // assume the frame length is always the same.
1428                         pts_int += length_per_frame;
1429                 }
1430         }
1431 }
1432
1433 void Mixer::render_one_frame(int64_t duration)
1434 {
1435         // Determine the time code for this frame before we start rendering.
1436         string timecode_text = timecode_renderer->get_timecode_text(double(pts_int) / TIMEBASE, frame_num);
1437         if (display_timecode_on_stdout) {
1438                 printf("Timecode: '%s'\n", timecode_text.c_str());
1439         }
1440
1441         // Update Y'CbCr settings for all cards.
1442         {
1443                 lock_guard<mutex> lock(card_mutex);
1444                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
1445                         YCbCrInterpretation *interpretation = &ycbcr_interpretation[card_index];
1446                         input_state.ycbcr_coefficients_auto[card_index] = interpretation->ycbcr_coefficients_auto;
1447                         input_state.ycbcr_coefficients[card_index] = interpretation->ycbcr_coefficients;
1448                         input_state.full_range[card_index] = interpretation->full_range;
1449                 }
1450         }
1451
1452         // Get the main chain from the theme, and set its state immediately.
1453         Theme::Chain theme_main_chain = theme->get_chain(0, pts(), global_flags.width, global_flags.height, input_state);
1454         EffectChain *chain = theme_main_chain.chain;
1455         theme_main_chain.setup_chain();
1456         //theme_main_chain.chain->enable_phase_timing(true);
1457
1458         // If HDMI/SDI output is active and the user has requested auto mode,
1459         // its mode overrides the existing Y'CbCr setting for the chain.
1460         YCbCrLumaCoefficients ycbcr_output_coefficients;
1461         if (global_flags.ycbcr_auto_coefficients && output_card_index != -1) {
1462                 ycbcr_output_coefficients = cards[output_card_index].output->preferred_ycbcr_coefficients();
1463         } else {
1464                 ycbcr_output_coefficients = global_flags.ycbcr_rec709_coefficients ? YCBCR_REC_709 : YCBCR_REC_601;
1465         }
1466
1467         // TODO: Reduce the duplication against theme.cpp.
1468         YCbCrFormat output_ycbcr_format;
1469         output_ycbcr_format.chroma_subsampling_x = 1;
1470         output_ycbcr_format.chroma_subsampling_y = 1;
1471         output_ycbcr_format.luma_coefficients = ycbcr_output_coefficients;
1472         output_ycbcr_format.full_range = false;
1473         output_ycbcr_format.num_levels = 1 << global_flags.x264_bit_depth;
1474         chain->change_ycbcr_output_format(output_ycbcr_format);
1475
1476         // Render main chain. If we're using zerocopy Quick Sync encoding
1477         // (the default case), we take an extra copy of the created outputs,
1478         // so that we can display it back to the screen later (it's less memory
1479         // bandwidth than writing and reading back an RGBA texture, even at 16-bit).
1480         // Ideally, we'd like to avoid taking copies and just use the main textures
1481         // for display as well, but they're just views into VA-API memory and must be
1482         // unmapped during encoding, so we can't use them for display, unfortunately.
1483         GLuint y_tex, cbcr_full_tex, cbcr_tex;
1484         GLuint y_copy_tex, cbcr_copy_tex = 0;
1485         GLuint y_display_tex, cbcr_display_tex;
1486         GLenum y_type = (global_flags.x264_bit_depth > 8) ? GL_R16 : GL_R8;
1487         GLenum cbcr_type = (global_flags.x264_bit_depth > 8) ? GL_RG16 : GL_RG8;
1488         const bool is_zerocopy = video_encoder->is_zerocopy();
1489         if (is_zerocopy) {
1490                 cbcr_full_tex = resource_pool->create_2d_texture(cbcr_type, global_flags.width, global_flags.height);
1491                 y_copy_tex = resource_pool->create_2d_texture(y_type, global_flags.width, global_flags.height);
1492                 cbcr_copy_tex = resource_pool->create_2d_texture(cbcr_type, global_flags.width / 2, global_flags.height / 2);
1493
1494                 y_display_tex = y_copy_tex;
1495                 cbcr_display_tex = cbcr_copy_tex;
1496
1497                 // y_tex and cbcr_tex will be given by VideoEncoder.
1498         } else {
1499                 cbcr_full_tex = resource_pool->create_2d_texture(cbcr_type, global_flags.width, global_flags.height);
1500                 y_tex = resource_pool->create_2d_texture(y_type, global_flags.width, global_flags.height);
1501                 cbcr_tex = resource_pool->create_2d_texture(cbcr_type, global_flags.width / 2, global_flags.height / 2);
1502
1503                 y_display_tex = y_tex;
1504                 cbcr_display_tex = cbcr_tex;
1505         }
1506
1507         const int64_t av_delay = lrint(global_flags.audio_queue_length_ms * 0.001 * TIMEBASE);  // Corresponds to the delay in ResamplingQueue.
1508         bool got_frame = video_encoder->begin_frame(pts_int + av_delay, duration, ycbcr_output_coefficients, theme_main_chain.input_frames, &y_tex, &cbcr_tex);
1509         assert(got_frame);
1510
1511         GLuint fbo;
1512         if (is_zerocopy) {
1513                 fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, y_copy_tex);
1514         } else {
1515                 fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex);
1516         }
1517         check_error();
1518         chain->render_to_fbo(fbo, global_flags.width, global_flags.height);
1519
1520         if (display_timecode_in_stream) {
1521                 // Render the timecode on top.
1522                 timecode_renderer->render_timecode(fbo, timecode_text);
1523         }
1524
1525         resource_pool->release_fbo(fbo);
1526
1527         if (is_zerocopy) {
1528                 chroma_subsampler->subsample_chroma(cbcr_full_tex, global_flags.width, global_flags.height, cbcr_tex, cbcr_copy_tex);
1529         } else {
1530                 chroma_subsampler->subsample_chroma(cbcr_full_tex, global_flags.width, global_flags.height, cbcr_tex);
1531         }
1532         if (output_card_index != -1) {
1533                 cards[output_card_index].output->send_frame(y_tex, cbcr_full_tex, ycbcr_output_coefficients, theme_main_chain.input_frames, pts_int, duration);
1534         }
1535         resource_pool->release_2d_texture(cbcr_full_tex);
1536
1537         // Set the right state for the Y' and CbCr textures we use for display.
1538         glBindFramebuffer(GL_FRAMEBUFFER, 0);
1539         glBindTexture(GL_TEXTURE_2D, y_display_tex);
1540         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
1541         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
1542         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
1543
1544         glBindTexture(GL_TEXTURE_2D, cbcr_display_tex);
1545         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
1546         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
1547         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
1548
1549         RefCountedGLsync fence = video_encoder->end_frame();
1550
1551         // The live frame pieces the Y'CbCr texture copies back into RGB and displays them.
1552         // It owns y_display_tex and cbcr_display_tex now (whichever textures they are).
1553         DisplayFrame live_frame;
1554         live_frame.chain = display_chain.get();
1555         live_frame.setup_chain = [this, y_display_tex, cbcr_display_tex]{
1556                 display_input->set_texture_num(0, y_display_tex);
1557                 display_input->set_texture_num(1, cbcr_display_tex);
1558         };
1559         live_frame.ready_fence = fence;
1560         live_frame.input_frames = {};
1561         live_frame.temp_textures = { y_display_tex, cbcr_display_tex };
1562         output_channel[OUTPUT_LIVE].output_frame(move(live_frame));
1563
1564         // Set up preview and any additional channels.
1565         for (int i = 1; i < theme->get_num_channels() + 2; ++i) {
1566                 DisplayFrame display_frame;
1567                 Theme::Chain chain = theme->get_chain(i, pts(), global_flags.width, global_flags.height, input_state);  // FIXME: dimensions
1568                 display_frame.chain = move(chain.chain);
1569                 display_frame.setup_chain = move(chain.setup_chain);
1570                 display_frame.ready_fence = fence;
1571                 display_frame.input_frames = move(chain.input_frames);
1572                 display_frame.temp_textures = {};
1573                 output_channel[i].output_frame(move(display_frame));
1574         }
1575 }
1576
1577 void Mixer::audio_thread_func()
1578 {
1579         pthread_setname_np(pthread_self(), "Mixer_Audio");
1580
1581         while (!should_quit) {
1582                 AudioTask task;
1583
1584                 {
1585                         unique_lock<mutex> lock(audio_mutex);
1586                         audio_task_queue_changed.wait(lock, [this]{ return should_quit || !audio_task_queue.empty(); });
1587                         if (should_quit) {
1588                                 return;
1589                         }
1590                         task = audio_task_queue.front();
1591                         audio_task_queue.pop();
1592                 }
1593
1594                 ResamplingQueue::RateAdjustmentPolicy rate_adjustment_policy =
1595                         task.adjust_rate ? ResamplingQueue::ADJUST_RATE : ResamplingQueue::DO_NOT_ADJUST_RATE;
1596                 vector<float> samples_out = audio_mixer->get_output(
1597                         task.frame_timestamp,
1598                         task.num_samples,
1599                         rate_adjustment_policy);
1600
1601                 // Send the samples to the sound card, then add them to the output.
1602                 if (alsa) {
1603                         alsa->write(samples_out);
1604                 }
1605                 if (output_card_index != -1) {
1606                         const int64_t av_delay = lrint(global_flags.audio_queue_length_ms * 0.001 * TIMEBASE);  // Corresponds to the delay in ResamplingQueue.
1607                         cards[output_card_index].output->send_audio(task.pts_int + av_delay, samples_out);
1608                 }
1609                 video_encoder->add_audio(task.pts_int, move(samples_out));
1610         }
1611 }
1612
1613 void Mixer::release_display_frame(DisplayFrame *frame)
1614 {
1615         for (GLuint texnum : frame->temp_textures) {
1616                 resource_pool->release_2d_texture(texnum);
1617         }
1618         frame->temp_textures.clear();
1619         frame->ready_fence.reset();
1620         frame->input_frames.clear();
1621 }
1622
1623 void Mixer::start()
1624 {
1625         mixer_thread = thread(&Mixer::thread_func, this);
1626         audio_thread = thread(&Mixer::audio_thread_func, this);
1627 }
1628
1629 void Mixer::quit()
1630 {
1631         should_quit = true;
1632         audio_task_queue_changed.notify_one();
1633         mixer_thread.join();
1634         audio_thread.join();
1635 }
1636
1637 void Mixer::transition_clicked(int transition_num)
1638 {
1639         theme->transition_clicked(transition_num, pts());
1640 }
1641
1642 void Mixer::channel_clicked(int preview_num)
1643 {
1644         theme->channel_clicked(preview_num);
1645 }
1646
1647 YCbCrInterpretation Mixer::get_input_ycbcr_interpretation(unsigned card_index) const
1648 {
1649         lock_guard<mutex> lock(card_mutex);
1650         return ycbcr_interpretation[card_index];
1651 }
1652
1653 void Mixer::set_input_ycbcr_interpretation(unsigned card_index, const YCbCrInterpretation &interpretation)
1654 {
1655         lock_guard<mutex> lock(card_mutex);
1656         ycbcr_interpretation[card_index] = interpretation;
1657 }
1658
1659 void Mixer::start_mode_scanning(unsigned card_index)
1660 {
1661         assert(card_index < num_cards);
1662         if (is_mode_scanning[card_index]) {
1663                 return;
1664         }
1665         is_mode_scanning[card_index] = true;
1666         mode_scanlist[card_index].clear();
1667         for (const auto &mode : cards[card_index].capture->get_available_video_modes()) {
1668                 mode_scanlist[card_index].push_back(mode.first);
1669         }
1670         assert(!mode_scanlist[card_index].empty());
1671         mode_scanlist_index[card_index] = 0;
1672         cards[card_index].capture->set_video_mode(mode_scanlist[card_index][0]);
1673         last_mode_scan_change[card_index] = steady_clock::now();
1674 }
1675
1676 map<uint32_t, VideoMode> Mixer::get_available_output_video_modes() const
1677 {
1678         assert(desired_output_card_index != -1);
1679         lock_guard<mutex> lock(card_mutex);
1680         return cards[desired_output_card_index].output->get_available_video_modes();
1681 }
1682
1683 string Mixer::get_ffmpeg_filename(unsigned card_index) const
1684 {
1685         assert(card_index >= num_cards && card_index < num_cards + num_video_inputs);
1686         return ((FFmpegCapture *)(cards[card_index].capture.get()))->get_filename();
1687 }
1688
1689 void Mixer::set_ffmpeg_filename(unsigned card_index, const string &filename) {
1690         assert(card_index >= num_cards && card_index < num_cards + num_video_inputs);
1691         ((FFmpegCapture *)(cards[card_index].capture.get()))->change_filename(filename);
1692 }
1693
1694 void Mixer::wait_for_next_frame()
1695 {
1696         unique_lock<mutex> lock(frame_num_mutex);
1697         unsigned old_frame_num = frame_num;
1698         frame_num_updated.wait_for(lock, seconds(1),  // Timeout is just in case.
1699                 [old_frame_num, this]{ return this->frame_num > old_frame_num; });
1700 }
1701
1702 Mixer::OutputChannel::~OutputChannel()
1703 {
1704         if (has_current_frame) {
1705                 parent->release_display_frame(&current_frame);
1706         }
1707         if (has_ready_frame) {
1708                 parent->release_display_frame(&ready_frame);
1709         }
1710 }
1711
1712 void Mixer::OutputChannel::output_frame(DisplayFrame &&frame)
1713 {
1714         // Store this frame for display. Remove the ready frame if any
1715         // (it was seemingly never used).
1716         {
1717                 lock_guard<mutex> lock(frame_mutex);
1718                 if (has_ready_frame) {
1719                         parent->release_display_frame(&ready_frame);
1720                 }
1721                 ready_frame = move(frame);
1722                 has_ready_frame = true;
1723
1724                 // Call the callbacks under the mutex (they should be short),
1725                 // so that we don't race against a callback removal.
1726                 for (const auto &key_and_callback : new_frame_ready_callbacks) {
1727                         key_and_callback.second();
1728                 }
1729         }
1730
1731         // Reduce the number of callbacks by filtering duplicates. The reason
1732         // why we bother doing this is that Qt seemingly can get into a state
1733         // where its builds up an essentially unbounded queue of signals,
1734         // consuming more and more memory, and there's no good way of collapsing
1735         // user-defined signals or limiting the length of the queue.
1736         if (transition_names_updated_callback) {
1737                 vector<string> transition_names = global_mixer->get_transition_names();
1738                 bool changed = false;
1739                 if (transition_names.size() != last_transition_names.size()) {
1740                         changed = true;
1741                 } else {
1742                         for (unsigned i = 0; i < transition_names.size(); ++i) {
1743                                 if (transition_names[i] != last_transition_names[i]) {
1744                                         changed = true;
1745                                         break;
1746                                 }
1747                         }
1748                 }
1749                 if (changed) {
1750                         transition_names_updated_callback(transition_names);
1751                         last_transition_names = transition_names;
1752                 }
1753         }
1754         if (name_updated_callback) {
1755                 string name = global_mixer->get_channel_name(channel);
1756                 if (name != last_name) {
1757                         name_updated_callback(name);
1758                         last_name = name;
1759                 }
1760         }
1761         if (color_updated_callback) {
1762                 string color = global_mixer->get_channel_color(channel);
1763                 if (color != last_color) {
1764                         color_updated_callback(color);
1765                         last_color = color;
1766                 }
1767         }
1768 }
1769
1770 bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
1771 {
1772         lock_guard<mutex> lock(frame_mutex);
1773         if (!has_current_frame && !has_ready_frame) {
1774                 return false;
1775         }
1776
1777         if (has_current_frame && has_ready_frame) {
1778                 // We have a new ready frame. Toss the current one.
1779                 parent->release_display_frame(&current_frame);
1780                 has_current_frame = false;
1781         }
1782         if (has_ready_frame) {
1783                 assert(!has_current_frame);
1784                 current_frame = move(ready_frame);
1785                 ready_frame.ready_fence.reset();  // Drop the refcount.
1786                 ready_frame.input_frames.clear();  // Drop the refcounts.
1787                 has_current_frame = true;
1788                 has_ready_frame = false;
1789         }
1790
1791         *frame = current_frame;
1792         return true;
1793 }
1794
1795 void Mixer::OutputChannel::add_frame_ready_callback(void *key, Mixer::new_frame_ready_callback_t callback)
1796 {
1797         lock_guard<mutex> lock(frame_mutex);
1798         new_frame_ready_callbacks[key] = callback;
1799 }
1800
1801 void Mixer::OutputChannel::remove_frame_ready_callback(void *key)
1802 {
1803         lock_guard<mutex> lock(frame_mutex);
1804         new_frame_ready_callbacks.erase(key);
1805 }
1806
1807 void Mixer::OutputChannel::set_transition_names_updated_callback(Mixer::transition_names_updated_callback_t callback)
1808 {
1809         transition_names_updated_callback = callback;
1810 }
1811
1812 void Mixer::OutputChannel::set_name_updated_callback(Mixer::name_updated_callback_t callback)
1813 {
1814         name_updated_callback = callback;
1815 }
1816
1817 void Mixer::OutputChannel::set_color_updated_callback(Mixer::color_updated_callback_t callback)
1818 {
1819         color_updated_callback = callback;
1820 }
1821
1822 mutex RefCountedGLsync::fence_lock;